Repository navigation
[Fix] The app may show raw <think> tags: frames are not split on the exclusive path #112
Description
The Kotlin client matches == "<think>" and == "</think>" exactly (ChatWebSocketClient.kt ~187, ~194). The exclusive path streams through TextIteratorStreamer, which emits at word/newline boundaries, so Qwen3's <think>\n and </think>\n\n probably arrive as one frame with the newline; adaptive_special_tag_buffering (models/base.py) joins but never splits. Then the reasoning panel never opens and the tags show up in the answer. The fake engines in the tests emit one token per chunk, so this is untested. Needs a real-model confirmation.
Fix: split tag frames on the server (each tag its own frame), and/or make the client tag-aware.
Done when: a test through /api/chat/streaming with a template that emits think tags, on the exclusive path, sees frames exactly <think> and </think>.
Found in the 2026-10-03 usability audit (code reading, not run).
Activity
b-re-w commentedon Oct 3, 2026
Confirmed without running a model: feeding Qwen3-0.6B's tokenizer output for
<think>\n…</think>\n\nHellotoken by token intoTextIteratorStreamer(the one-at-a-time path) yields the frames'<think>\n'and'</think>\n\n', never an exact<think>/</think>, so the app's exact match fails. Since #127 default requests batch, but requests withrepeat_penalty/typical_p/extra kwargs still take this path. Fix: split tags into their own frames on the server.b-re-w commentedon Oct 3, 2026
Landed in develop via #136